Can a machine write a poem that moves you? It's one of the oldest questions in AI, and in 2026 it finally has a genuinely interesting answer.
We ran an experiment. We gave three AI models — GPT, Claude, and Gemini — the same five poetry prompts. We also gave the same prompts to a Pulitzer-nominated poet. Then we ran every poem through our AI detector and asked a panel of 50 readers to rate each poem for beauty, originality, and emotional impact.
Here's what happened.
The AI poems were, honestly, not bad. They had rhythm, imagery, and emotional language. GPT's "last autumn leaf" poem opened with a striking line: "She held on with the stubbornness of a name you can't remember." That's genuinely good.
But the AI poems shared a pattern: they were uniformly polished. Every stanza was well-crafted. Every metaphor landed. There were no missteps — and that's exactly the problem. Real human poetry has rough edges. It takes risks that don't always pay off. The AI never took a risk it couldn't make work.
The Pulitzer-nominated poet's work was messier, rawer, and more memorable. On the prompt "a conversation you never had," the human wrote:
> *I keep almost calling you, / and the almost is louder / than anything I'd have said.*
One reader said: "I felt that in my chest." No one said that about an AI poem.
Our AI detector correctly flagged all 15 AI poems as AI-generated with scores above 90%. The human poems scored below 15%. The detector didn't even struggle. Why? Because the statistical fingerprints of AI — consistent tone, predictable transitions, and the absence of true unpredictability — show up clearly in poetry. Poetry is the genre where burstiness matters most, and AI is consistently low in burstiness.
Can AI write better poetry than humans? On technique alone, the gap is closing. On emotional impact, humans still win easily. The poet's work had specificity — a real memory, a real ache. The AI's work had the shape of emotion without the weight of it.
There's a beautiful irony here: the very thing that makes human writing imperfect — its unevenness, its unpredictability — is also what makes it real. And it's what our detector is built to find.